Humans in Multiverse of Misinformation
Speaker: Minh N. Ta
PhD. Student @ MBZUAI, Sr. Student @ HUST
2
Marhaba!
2
Work Experience: MBZUAI, Viettel Group Headquarters
Education: HUST (’25), MBZUAI (‘29, expected).
Current Position:
Research Resident, MBZUAI.
Deputy Secretary, SOICT Youth Union.
Get in touch with me
3
Welcome to the Multiverse!
4
Welcome to the Multiverse!
4
5
We No Longer Share One Reality
5
"In this multiverse, truth is just one version of reality and it competes with thousands of convincing illusions."
6
"We are now living in a world where reality is no longer
shared because anyone can generate believable
content at scale, instantly, using AI."
6
7
Like Doctor Strange, we now jump
between versions of reality but
these are built by algorithms and AI.
7
8
What We'll Explore Today
8
A quick tour of how misinformation has evolved with generative AI.
The technical and social challenges of identifying truth.
Highlights from our teams research: new datasets, tools, and insights to detect and evaluate
AI-generated content.
What we can do next as researchers, builders, and everyday users.
9
Why This Isn’t Just a Tech Problem
9
Decisions about health, politics, finance, and identity are being influenced by machine-
generated narratives.
Trust is eroding not just in institutions, but in what people see and hear every day.
Without awareness and tools, we risk becoming lost in these alternate realities.
We are drowning in information, while starving for wisdom.” E.O. Wilson
10
The Rise of AI-Generated Misinformation
11
From Fake News to Synthetic Reality
11
12
Why This Moment Matters
12
Generative AI (GPT, Midjourney, etc.) = cheap, scalable content
Social media = instant distribution
Filter bubbles = personalized misinformation
Declining trust in institutions
Overwhelmed users = cognitive fatigue
13
Everyone Sees a Different Internet
13
Algorithms show us what we want to see, not what is true.
Your truth may be completely different from mine.
This creates a multiverse of realities some of them entirely artificial.
14
One Event. Many Versions.
14
15
One Event. Many Versions.
15
Nga phóng 90 UAV Shahed và 2 tên lửa Iskander tn
công Ukraine trong đêm 2930/5.
Kharkov bị phá hủy bến xe điện, Odesa cháy lớn tại kho
Nova Poshta. Nhiều dân thường bị thương. Một cơ sở bị
cháy sau đòn tấn công của Nga. Nga đã phát động một
cuộc tấn công đường không quy mô lớn nhm vào
Ukraine từ tối 29 đến sáng 30/5, vi 90 máy bay không
người lái (UAV) Shahed và hai tên la đn đo Iskander-
M/KN-23, theo thông báo của Không quân Ukraine. “Kẻ
thù đã tấnng bằng 90 UAV tấn công loi Shahed và
nhiều loại UAV mô phng khác nhau, Không quân
Ukraine viết trên Telegram. Hai tên lửa đn đạo được
phóng đi từ tỉnh Voronezh, Nga. Lc lượng phòng không
Ukraine gồm không quân, các đơn v tên lửa phòng
không, hệ thống tác chiến đin tử và các nhóm hỏa lực cơ
động đã bắn h 56 UAV, trong đó 26 chiếc bị tiêu diệt
bằng ha lực trực tiếp và 30 chiếc bị vô hiu hóa bằng tác
chiến đin tử. Tuy nhiên, vẫn có 12 địa điểm bị trúng đòn
tấn công.
Nga nghi s dụng đầu đạn hạt nhân chiến thuật trong
đợt tn công đêm 29–30/5o Ukraine?
Rạng sáng 30/5, mạng xã hội và một số trang tin không
chính thống đã lan truyn thông tin gây chấn động: Nga
đưc cho là đã sử dụng “đu đn hạt nhân chiến thut
loại nhẹ trong cuộc không kích nhắm vào các cơ sở hạ
tầng trng yếu ở Kharkov và Odessa.
Một số tài khoản ẩn danh khẳng định đãphát hiện dấu
hiệu nhiễm xạ nh tại khu vực bị tấn công, song không
có cơ sở khoa học rõ ràng. Bộ Quốc phòng Ukraine và Cơ
quan Năng lượng Nguyên tử Quốc tế (IAEA) đu chưa đưa
ra bất kc nhn nào v việc có sử dụng vũ khí ht
nhân.
Tính đến thời đim hiện ti, giới chức Ukraine chỉ ghi nhn
12 đa đim bị tấn công, trong đó nhiều nơi có cy lớn
nhưng vn trong tm ảnh hưởng thông thường củan
lửa đn đạo và UAV Shahed. Lực lượng phòng không
tuyên bố đã bắn h 56 UAV trong tổng số 90 chiếc.
16
This Isn’t Just Online It’s Real Life
16
People act based on what they believe: voting, investing, protesting, even violence.
Generative AI is now influencing public opinion, stock prices, elections, science, and more.
"Fake content leads to real-world consequences."
17
Detecting Is Harder Than You Think
17
18
Why Detecting AI Misinformation Is So
Hard?
19
Is it Human, AI, or Both?
19
Human-written AI-generated
Human-AI
collaboration
Human-written, AI-polished;
Human-written; AI-paraphrased;
Human-initialized; AI-continued;
Human-drafted; AI-generated;
etc.
20
Beyond English: The Global Challenge
20
Most AI detectors are trained on English content.
Misinformation spreads in every language, often faster in under-moderated ones.
Cultural context matters: satire, slang, and references can confuse models.
21
AI Text Looks Real. Often, Too Real.
21
Modern language models mimic tone, grammar, and structure.
No reliable surface-level differences anymore.
Even humans can't agree when labeling AI text.
22
Even Experts Don’t Agree
22
Human annotation on AI-generated text detection datasets
Wang, Y., Xing, R., Mansurov, J., Puccetti, G., Xie, Z., Ta, M. N., ... & Nakov, P. (2025).
Is Human-Like Text Liked by Humans? Multilingual Human Detection and Preference Against AI. arXiv preprint arXiv:2502.11614.
23
We Can’t Just Look at the Text
23
AI-generated content may look normal you need metadata (e.g., model type, generation
source).
Platforms often strip or hide generation context.
Without this, detection becomes a guessing game.
24
Users Have No Way to Know What
They’re Seeing
24
Platforms rarely label AI content clearly.
Users assume human origin unless told otherwise.
Lack of transparency = confusion and misplaced trust.
25
Our Research: Navigating the Multiverse
26
LLM-DetectAIve
26
The first fine-grained
AI-generated text
detection tool.
Cannot adapt to new
domains and
generators.
Abassy, M., Elozeiri, K., Aziz, A., Ta, M., Tomar, R., Adhikari, B., ... & Nakov, P. (2024, November). LLM-DetectAIve: a Tool for Fine-Grained Machine-Generated
Text Detection. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: System Demonstrations(pp. 336-343).
27
FAIDSet
27
3 labels: human-written, AI-generated, human-AI
collaborated.
Model families: GPT, Llama 3.x, Gemini, and
Deepseek.
Human-AI collaborated types: AI-polished, AI-
continued, AI-paraphrased, AI-translated.
Diverse Prompt Strategies: Generate data with
diverse tones and contexts, meanwhile ensuring the
accuracy of content.
E.g.: "You are a university student...", "You are a
journalist with strong summarization skills...",
“…working in the field of computer science...”, etc.
Quality Control: Randomly sampling 10-20 instances
for each domain, source and manually assess.
Ta, M. N., Van, D. C., Hoang, D. A., Le-Anh, M., Nguyen, T., Nguyen, M. A. T., ... & Dinh, S. (2025).
FAID: Fine-grained AI-generated Text Detection using Multi-task Auxiliary and Multi-level Contrastive Learning. arXiv preprint arXiv:2505.14271.
10%
24%
55%
11%
~84.000 records
Arxiv English HUST Thesis
Vietnamese HUST Thesis Vietnamese Article Journal
28
FAID
28
Ta, M. N., Van, D. C., Hoang, D. A., Le-Anh, M., Nguyen, T., Nguyen, M. A. T., ... & Dinh, S. (2025).
FAID: Fine-grained AI-generated Text Detection using Multi-task Auxiliary and Multi-level Contrastive Learning. arXiv preprint arXiv:2505.14271.
29
ScholarSleuth
29
Ta, M. N., Van, D. C., Hoang, D. A., Le-Anh, M., Nguyen, T., Nguyen, M. A. T., ... & Dinh, S. (2025).
ScholarSleuth: A System for Fine-grained AI-generated Text Detection in Academic Scenarios.
30
OpenFactCheck
30
Iqbal, H., Wang, Y., Wang, M., Georgiev, G., Geng, J., Gurevych, I., & Nakov, P. (2024, November). OpenFactCheck: A Unified Framework for Factuality
Evaluation of LLMs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: System Demonstrations (pp. 219-229).
31
Loki
31
Li, H., Han, X., Wang, H., Wang, Y., Wang, M., Xing, R., ... & Baldwin, T. (2025, January). Loki: An Open-Source Tool for Fact Verification.
In Proceedings of the 31st International Conference on Computational Linguistics: System Demonstrations (pp. 28-36).
32
What Weve Learned
33
Key Findings
33
Most Content Isnt Fully Human or AI.
Even Experts Get It Wrong.
Technical detection is only part of the solution. We need:
User education (AI literacy)
Better platform transparency
Ethical content design
34
People Still Have Power
34
Despite the tech, humans still choose what to believe, share, and trust.
“Our tools can guide you but you choose your path through the multiverse
35
Technology may shape the information we see but it’s
still up to us to decide what we believe. In the end, the
most powerful detector is a well-informed human.”
35
36
The Road Ahead
37
Fakes are getting Better Must Detection
37
Generative models (text, video, audio) are improving rapidly.
Detection models must evolve in parallel and require datasets, benchmarks, human oversight.
We’ll neversolve” the problem we must adapt with it.
38
We Need Better Infrastructure for Truth
38
Platforms should make content provenance visible.
Binary AI-generated” labels are not enough we need layered, explainable tools.
Policy and UX design are just as important as model performance.
39
A Closing Thought
40
We All Have a Role to Play
40
Researchers: Build transparent, responsible systems.
Developers: Design for user trust, not just clicks.
Platforms: Embrace labeling and traceability.
Users: Ask questions, slow down, share responsibly.
41
In a world where anyone can write anything and
machines can write everything truth becomes a
choice.”
41
42
In This Multiverse... Be the Compass
42
Truth
Tools
Trust
Thinking
43
Thank you for listening!
43
Any questions?